IEEE TRANSACTIONS ON VISUALIZATION AND COMPUTER GRAPHICS 1

Towards Visualization Thumbnail Designs that

Entice Reading Data-driven Articles

Hwiyeon Kim, Joohee Kim, Yunha Han, Hwajung Hong, Oh-Sang Kwon, Young-Woo Park,

Niklas Elmqvist, Senior Member, IEEE, Sungahn Ko, and Bum Chul Kwon

Abstract—As online news increasingly include data journalism, there is a corresponding increase in the incorporation of visualization

in article thumbnail images. However, little research exists on the design rationale for visualization thumbnails, such as resizing,

cropping, simplifying, and embellishing charts that appear within the body of the associated article. Therefore, in this paper we aim to

understand these design choices and determine what makes a visualization thumbnail inviting and interpretable. To this end, we ﬁrst

survey visualization thumbnails collected online and discuss visualization thumbnail practices with data journalists and news graphics

designers. Based on the survey and discussion results, we then deﬁne a design space for visualization thumbnails and conduct a user

study with four types of visualization thumbnails derived from the design space. The study results indicate that different chart

components play different roles in attracting reader attention and enhancing reader understandability of the visualization thumbnails.

We also ﬁnd various thumbnail design strategies for effectively combining the charts’ components, such as a data summary with

highlights and data labels, and a visual legend with text labels and Human Recognizable Objects (HROs), into thumbnails. Ultimately,

we distill our ﬁndings into design implications that allow effective visualization thumbnail designs for data-rich news articles. Our work

can thus be seen as a ﬁrst step toward providing structured guidance on how to design compelling thumbnails for data stories.

Index Terms—Data journalism, data-driven storytelling, online news, visualization, thumbnail images, data stories.

✦

1 INTRODUCTION

HUMBNAILS are small and static images that accom-

pany titles and bylines in documents [1], [2], [3]. News

organizations have extensively used thumbnails for their

online news articles to induce reader attention and attract

clicks. As data stories—where visualization plays an im-

portant role—grow in popularity, news organizations are

starting to publish thumbnails that incorporate visualization

into the thumbnails. These visualization thumbnails are differ-

ent from conventional thumbnails because they have great

potential to both convey an article’s main message using its

embedded visualization as well as entice the reader to click

on the article link based on the appeal of the displayed data.

Several news media outlets have begun to adopt such

visualization thumbnails. For example, The New York Times

presents many visualization thumbnails in its data story

dedicated section “The Upshot." FiveThirtyEight showcases

many visualization thumbnails that can effectively catch the

reader’s attention. Although many types of visualization

thumbnails have been designed, few studies have inves-

tigated the variety of design choices. As a result, many

questions remain unanswered about optimal visualization

thumbnail design. For example, “The Upshot” and The

Economist tend to use thumbnails that are resized from

visualizations embedded in the article. Meanwhile, The Pud-

ding, FiveThirtyEight, and The Wall Street Journal often design

visualization thumbnails with editing choices different from

• H. Kim, J. Kim, Y. Han, O. Kwon, Y. Park, and S. Ko (corresponding

author) are with UNIST, Ulsan, South Korea.

• H. Hong is with KAIST, Daejeon, South Korea.

• N. Elmqvist is with the University of Maryland, College Park, MD, USA

• B.C. Kwon is with IBM Research, Cambridge, MA, USA

Manuscript received XX YY, 2020; revised AA BB, 2021.

the source visualizations in the article by adding annotations

or removing axes. However, it is unclear which factors

inﬂuence both interpretability—helping readers understand

the article from the thumbnail alone—and appeal—inducing

the reader to click to read the article.

In this research, we aim to investigate visualization

thumbnail designs that support both of these affordances.

To this end, we ﬁrst survey existing visualization thumb-

nails from news organizations and ﬁnd different types of

thumbnail designs (Section 3.1). Then we interview six

industry practitioners to understand their thumbnail design

strategies and intentions (Section 3.2). With the survey and

interview results, we conduct a user study (n = 161)

using 16 thumbnails that we carefully design with industry

practitioners to investigate (RQ1) visualization thumbnail

designs that readers want to see, and (RQ2) the roles of

thumbnail components (Section 4). The results indicate that

readers want visualization thumbnails that quickly capture

their attention and effectively convey the articles’ main

point (Section 5). We report the lessons learned and design

implications derived from the study and discuss open re-

search areas on visualization thumbnails (Section 6.1). To

our knowledge, this is the ﬁrst attempt to explore the design

space, component roles, and readers’ choice of visualization

thumbnails with evaluation.

This paper is a signiﬁcantly extended version of a IEEE

VIS 2019 short paper surveying visualization thumbnails for

data stories [4]. Compared to this earlier report, this paper

includes empirical evaluation involving 161 crowdworkers

as well as our corresponding analysis and ﬁndings. The

supplemental material for this journal version of the work

can be found on OSF: https://osf.io/khgw2/

The main contributions of this work are as follows: (1)

This article has been accepted for publication in IEEE Transactions on Visualization and Computer Graphics. This is the author's version which has not been fully edited and

content may change prior to final publication. Citation information: DOI 10.1109/TVCG.2023.3278304

Authorized licensed use limited to: Hong Kong University of Science and Technology. Downloaded on July 19,2023 at 07:46:22 UTC from IEEE Xplore. Restrictions apply.

IEEE TRANSACTIONS ON VISUALIZATION AND COMPUTER GRAPHICS 2

Extracting the key components of visualization thumbnail

designs by surveying in-the-wild visualization thumbnails;

(2) identifying reader preferences for visualization thumb-

nail designs and their rationale for their preferences via

a crowdsourced user study; and (3) reporting lessons, de-

sign implications, and open research areas for visualization

thumbnails.

2 RELATED WORK

We situate our work in relation to thumbnail design and to

the use of visualization for communication, particularly in

data journalism.

2.1 Thumbnail Design

Prior research has shown that thumbnails help people locate

and rediscover content during web searches, media brows-

ing, and sensemaking. For example, when searching the

web, thumbnails paired with titles, text snippets, and URLs

help people ﬁnd articles of interest online [2], [5], [6]. Aula

et al. [2] show that image thumbnails add information

about the relevance of the web page compared to the text

summary. Thumbnails also appear when browsing other

types of media such as ﬁle systems [7], documents [8],

and videos [9]. When an email contains a link to a video

ﬁle, Topkara et al. [10] ﬁnds that more people click on the

link displayed with the thumbnail. Previous work on visu-

alization and visual analysis has incorporated thumbnails

into the sensemaking process as a means of leveraging the

unique advantages of spatial memory [11], [12], [13].

In this study, we investigate the role thumbnails play

as people browse online news articles in the context of

data journalism. On online news webpages, thumbnails of

articles must compete for the readers’ attention with one an-

other and with other content. There are several factors that

affect the interest and ultimate usefulness of a thumbnail,

such as the size of the thumbnail and the text it contains.

For example, Kaasten et al. [14] conclude that thumbnails

must be larger than 96 × 96 pixels to trigger recognition

among returning readers to pages with multiple thumbnails.

Woodruff et al. [6] also ﬁnd that enhanced versions of

thumbnails including keywords have better search perfor-

mance than text summaries or image thumbnails.

Previous studies have been conducted to automatically

generate image thumbnails by cropping, reducing, or se-

lecting salient parts of photography included in articles [1],

[15], [16]. Song et al. [3] present an algorithm for selecting

the most salient and evocative thumbnail among image

candidates to represent a video. Compared to this prior art,

our study is the ﬁrst to study visualization images, as well as

the ﬁrst step towards the automatic creation of visualization

thumbnails based on a design space for data journalism.

2.2 Storytelling in Thumbnails

Thumbnails can be a key player in visual communication

and storytelling [17], [18], as their innate goal is to effectively

convey a story in a concise visual format. But as they are

usually given a smaller space budget, there have been many

discussions on “how much is much” (i.e., a debate between

minimalism and chartjunk). Many visualization researchers

have conventionally supported minimalism in visualization,

stating “less is more,” as they believe that objective charac-

teristics in visualization improve visualization performance.

Edward Tufte [19], one of the most vocal proponents of

maximizing the “data-to-ink” ratio, argues that “ink that

fails to depict statistical information does not have much

interest to the viewer.”

There are researchers who argue against minimalism in

visualization, stating that complexity (e.g., chartjunk [20])

is not a bad thing and could make a design more attrac-

tive [21]. Many experiments have been performed over the

years, but results seem to support that “chart junk” or

embellished graphics (e.g., Nigel Holmes’ work) is not a bad

thing. Kelly [22] conducted such an experiment and found

that 120 participants made similar errors in comprehension

and search tasks with the infographics data (collected from

USA Today), regardless of data-ink ratios. Other researchers

also observe in their experiments that visual embellishment

does not have a signiﬁcant impact on visualization perfor-

mance [23], [24]. Rather, visual embellishment even results

in better results in supporting human recall ability [20], [25],

[26], [27] and qualitative dimensions (e.g., aesthetics and

preference) [28], [29]. Due to the importance of annotations

(i.e., explanation text in this work) in story-telling, which is

also shown in our study, Ren et al. [30] propose a visual tool

for efﬁciently producing annotations, which can be greatly

useful for making thumbnails with ample context.

Annotations or other visual components may carry a bias

or slant [31]. Whether intended or not [32], a spotted bias

on thumbnail’s annotations negatively impacts an article’s

credibility and legitimacy. Kong and Agrawala [33] show

that placing additional layered information could aid chart

reading. We report how visual components on visualization

thumbnails can impact article reading and credibility.

2.3 Visualization in Data Journalism

With data journalism on the rise, visualization is becom-

ing pervasive in news media [34]. Different intentions for

visualization communication often lead to different design

choices [35], encountering substantial use of graphical and

text-based annotation [30]. We see that some visualizations

incorporate human-recognizable objects [36], such as icons

and logos, to help people understand the data. Other graph-

ics that are not directly relevant to data [37], such as portraits

and illustrations, are also incorporated into visualizations

and considered embellishments [20], [23] to enhance the

meaning of the chart. These graphical annotations are con-

sidered aesthetic design choices [24] that make positive

ﬁrst impressions [38] on the reader. However, there is still

debate as to whether these decorations have a positive effect

on the informativeness of charts. While some studies have

shown that embellishments can increase comprehension [33]

and memory [25], [26], [27], Hullman and Adar [18] argue

that some embellishments make charts more difﬁcult to

interpret. Andry et al. [39] quantitatively measure people’s

understanding of chart embellishments and show that there

is a positive effect within limited boundaries.

Since online news articles must hook users in a limited

time [40], the design of the ﬁrst impression [38] is vital.

Transparency of visualization is consistently emphasized

This article has been accepted for publication in IEEE Transactions on Visualization and Computer Graphics. This is the author's version which has not been fully edited and

content may change prior to final publication. Citation information: DOI 10.1109/TVCG.2023.3278304

Authorized licensed use limited to: Hong Kong University of Science and Technology. Downloaded on July 19,2023 at 07:46:22 UTC from IEEE Xplore. Restrictions apply.

IEEE TRANSACTIONS ON VISUALIZATION AND COMPUTER GRAPHICS 3

as a challenge in data journalism [41]. Some practices of

generating slanted titles [31] or designing deceptive visual-

izations [32] can bias and often mislead readers. Designing

thumbnails with visualizations is also important because the

thumbnails represent their articles and provide ﬁrst impres-

sions for the articles. Therefore, there is a need for research

on how to design attractive and effective thumbnails that

can be easily interpreted without misleading the reader.

Building on the above understanding, we further explore

the current practices of visualization thumbnail design.

3 VISUALIZATION THUMBNAIL DESIGN PRACTICE

To our knowledge, prior research has yet to examine the

design of thumbnails in data journalism, and particularly

those that we refer to as visualization thumbnails. To better

understand current practices in visualization thumbnail de-

sign, we collected examples from data journalism outlets

and conducted semi-structured interviews with six graphic

designers working for news websites.

3.1 A Survey of Visualization Thumbnails

We began by collecting news articles published between

November 1, 2018 and December 31, 2018 from online news

organizations reputed for their data journalism, the Pew Re-

search Center (Pew.), The Economist (Eco.), The New York Times

(NYT), including The Upshot and DealBook, FiveThirtyEight

(538), The Wall Street Journal (WSJ), First Tuesday Journal

(1st), and Bloomberg News (BBG). We concentrated on articles

relating to politics and economics, as these topics tend to

use visualization to a greater degree than others covered by

these news outlets. This initial corpus contained 139 articles

that include visualizations within the article body. Among

these, 48 articles did not feature visualizations in the article’s

thumbnail, instead opting for photographs and other visual

imagery aside from visualization. Of the remaining 91, we

decided to focus on basic charts, such as bar and line

charts and scatterplots, because they are most common in

news media. Thus we excluded 24 articles whose charts are

not axis-based (e.g., maps) or are infrequently used (e.g.,

pictogram charts, Sankey diagrams).

Among the remaining 67 articles, 39 included a visual-

ization used in the body of the article as its thumbnail, re-

ducing its size or cropping it. Thumbnails for the remaining

28 articles modiﬁed a visualization used in the body of the

article in some way (e.g., omitting axes). Examining these

67 thumbnails further, we codiﬁed 1) which visualization

components were modiﬁed (e.g., an axis); and 2) how com-

ponents were modiﬁed (e.g., omitted). Three authors of this

paper independently codiﬁed these components and later

merged their codes in an iterative discussion, arriving at

96% agreement (Fleiss’ Kappa = 0.75).

To label the components, we initially considered using

existing classiﬁcations, namely Borkin et al.’s classiﬁcation

of visualization components [27], Byrne et al.’s [37] distinc-

tion between graphical and ﬁgurative components [37], and

Ren et al.’s [30] classiﬁcation of annotation. We realized that

these classiﬁcations were insufﬁcient in isolation in terms

of capturing all aspects of visualization thumbnail design.

For example, Borkin et al.’s classiﬁcation deﬁnes “text” as

100b

80b

60b

40b

20b

03 05 07 09 11 13 15 17 19

Market capitalization (in billions)

Apple vs. Microsoft

Apple

Microsoft

$67b

the gap between companies

continued to grow

Apple

Microsoft

Chart Title

HRO

Data Label

Y Axis Label

X Axis

Highlight

Exp. text

Legend (Explicit)

Legend

(Implicit)

Y Axis Title

X Axis Title

Years

Fig. 1. Line chart annotated (in red) according to our classiﬁcation

of chart elements. The chart features examples of additional compo-

nents including HROs and highlights, as well as basic components such

as axes, data labels, and legends.

Visualization in an article

Resized

Cropped

Modied

Fig. 2. Generation strategies. Many news outlets generate thumbnails

by resizing, cropping, and modifying original visualizations in articles.

“any text in the image,” so axis titles, annotations, and

captions fall under the same category. Meanwhile, Byrne

et al.’s classiﬁcation provided broad categories for coding

ﬁgurative components. Finally, Ren et al.’s classiﬁcation

could not be used to describe chart components beyond an-

notation. While these classiﬁcations informed our analysis,

we struggled to use them as a means to codify the designers’

intentions or goals. We therefore combined and extended

the aforementioned classiﬁcations, resulting in a new classi-

ﬁcation having categories that explicitly acknowledge each

visualization element’s role.

As a result, we identiﬁed 14 basic and 4 [R2-18]additional

component types. The basic component types refer to chart

components that directly facilitate chart reading and under-

standing (e.g., two axes in the line chart). Figure 1 shows

examples of the basic component types. They include: x-

and y-axes with labels, tick marks, data labels, chart titles,

and legends. By data labels, we mean any text that directly

reﬂects a data value (e.g., ‘$67b’ in Figure 1). We further dis-

tinguish two types of legends: explicit legends: those drawn

in a dedicated area; and implicit legends: those drawn directly

within the visualizations (e.g., Apple & Microsoft in Fig-

ure 1). Additional components include explanatory text (or

exp. text), highlights (e.g., the vertical arrow in Figure 1, the

red bar in this thumbnail [42], Human Recognizable Objects

(HROs) [27]), and Graphics Not Relevant to Data (GNRD)

to capture all forms of graphical embellishment. HROs are

pictorial components used in legends (e.g., the Apple and

Microsoft logos in Figure 1 and a small human object in

this thumbnail [43] or to encode data points [36]. GNRDs

This article has been accepted for publication in IEEE Transactions on Visualization and Computer Graphics. This is the author's version which has not been fully edited and

content may change prior to final publication. Citation information: DOI 10.1109/TVCG.2023.3278304

Authorized licensed use limited to: Hong Kong University of Science and Technology. Downloaded on July 19,2023 at 07:46:22 UTC from IEEE Xplore. Restrictions apply.

IEEE TRANSACTIONS ON VISUALIZATION AND COMPUTER GRAPHICS 4

TABLE 1

Our classiﬁcation of 67 visualization thumbnails. Filled values reﬂect

modiﬁcations from a visualization found in the article. An interactive

version of this table can be accessed at

http://hci.unist.ac.kr/SurveyVisThumbnails/

Removed Added Remained Editing Type

538

Pew.

Eco.

NYT

WSJ

BBG

1st

538

Pew.

Eco.

NYT

SUM

Y Axis Label

Y Grid Line

X Axis Label

Chart Title

Legend Implicit

Y Axis Title

X Axis

X Tick Marks

Legend Explicit

X Grid Lines

Data Label

Y Axis

Y Tick Marks

X Axis Title

Exp. text

GNRD

HRO

Modified

Cropped

Resized

SUM

Highlight

Pict.

DP Scatterplot

Bubble

Bar + Line Bar Line

BASIC CHART COMP. ADDED CHART COMP.

41 31 29 27 26 24` 21 1214 10 9 2 36 36 6 3 2829 18

Fig.1&

Article 3

Fig.3

Article 1

Article 2

Fig.2-right

Fig.2-left

are images or illustrations that reﬂect the article’s context

but are not directly related the data, such as the blue image

(bottom-right) in this thumbnail [43]. Figure 2 describes the

current thumbnail creation strategies of news organizations

(e.g., resized, cropped, and modiﬁed thumbnails).

Table 1

presents the result of coding 67 visualization

1. Online version: http://bitly.kr/AGRYQA

TABLE 2

Thumbnail classiﬁcation. 24 visualization thumbnails that are not

axes-based (e.g., maps) or are infrequently used.

Modified

Cropped

Resized

Map Pie PCP San.

line+

map

bar+

line+

map

Rad.

BBG Pew. Eco. NYT WSJ FTJ NYT 538 Eco. Eco.BBG Pew.

SUM

thumbnails

using our basic and additional component clas-

siﬁcation, along with our modiﬁed / resized / cropped dis-

tinction; we also indicate chart types and sources along the

table’s vertical axis. ‘DP’ and ‘Pict.’ in the ﬁrst column (i.e.,

chart type) corresponding to dot plots and pictograms chart,

respectively.

Our codiﬁcation suggests several trends. Thumbnails for

line charts (30 out of 67) tend to omit the X-axis title, the Y-

axis, and legends. However, they tend to include additional

components such as highlights and explanation text. Bar

charts and scatterplots tend to include diverse combinations

of components that are omitted or added in thumbnails.

Example components used for the combinations include X-

axis grid lines, titles, and Y-axis. We are also able to contrast

the strategies of different news organizations. For instance,

FiveThirtyEight tends to remove nearly all components from

line charts in thumbnails; they also tend to add GNRDs

and HROs (e.g., [44]). Meanwhile, traditional print media

organizations such as The New York Times and The Economist

tend to crop or resize charts when producing thumbnails.

Many media organizations seem to prefer modiﬁcation (28)

and crop (29) strategies in their designs. In particular, we see

that FiveThirtyEight and the Pew Research Center are more

inclined to modify existing visualizations, while the New

York Times seems to prefer cropping existing visualizations

embedded in articles. Lastly, we see considerable variability

in Table 1 which is an indication of a need for greater

understanding in terms of how visualization thumbnail de-

sign choices affect readers’ interpretation as well as readers’

likelihood to read the article.

Table 2 shows the codiﬁcation results of the editing

strategies (i.e., modiﬁed, cropped, or resized) of 24 visu-

alization thumbnails, including maps, pie charts, parallel

coordinates, radial column charts (Rad.), Sankey diagrams

(San.), and composite charts. We coded the editing strategies

separately for the additional visualizations because those

strategies can be applied to any visualizations. As seen in

Table 2, cropping is more popular than the other two strate-

gies for the additional 24 visualizations; however, cropping

is not especially popular in the 67 visualizations, which

include X-axis and Y-axis, as shown in Table 1. Most maps

are presented as cropped, losing chart titles or legends. In

analyzing the thumbnails, such a removal seemed effective;

it allows more space without causing additional difﬁculty

in recognizing color-coded legends and small text. The

discussed trends in this section are examples and by no

means comprehensive.

2. Thumbnail collection: http://bitly.kr/bibRs4

This article has been accepted for publication in IEEE Transactions on Visualization and Computer Graphics. This is the author's version which has not been fully edited and

content may change prior to final publication. Citation information: DOI 10.1109/TVCG.2023.3278304

Authorized licensed use limited to: Hong Kong University of Science and Technology. Downloaded on July 19,2023 at 07:46:22 UTC from IEEE Xplore. Restrictions apply.

IEEE TRANSACTIONS ON VISUALIZATION AND COMPUTER GRAPHICS 5

3.2 Interviews with Visualization Designers

To understand the current practices in visualization thumb-

nail design further, we engaged in informal interviews with

six visualization designers from our extended professional

networks who are employed by news media organizations.

These included two journalist-engineers (with 6 and 16

years of experience, respectively), an interactive graphics

developer (with 15 years of experience), a senior news

article editor (with 12 years of experience), a data scientist

(with 6 years of experience), and a computational journalist

(with 3 years of experience). Given the demanding nature of

their work and the time zone differences between us, these

interviews were asynchronous and occurred via email or

within the #journalism channel of the Data Visualization

Society’s Slack workspace.

We asked these practitioners about two topics: (1) their

intentions with respect to designing thumbnails for articles

that prominently feature visualization; and (2) the chal-

lenges of incorporating visualization into article thumb-

nails. These practitioners reported a broad set of goals for

thumbnails. First, their thumbnails must build and reinforce

the organization’s brand identity. This requirement often

constrains the choice of colors, font types, and HROs used in

thumbnails. Second, their thumbnails must be aesthetically

appealing in order to draw readers’ attention, particularly

when the thumbnail appears in a visually rich social media

feed. Third, their thumbnails must also reﬂect any unique

artwork or visual content commissioned for the article (if

applicable), which might include illustrations, collages, or

animations.

In regards to how they design thumbnails for data

stories, these practitioners indicated that there are “no hard

and fast rules of thumb” that can be applied to all cases.

Aside from using a limited color palette to reinforce brand

recognition, two of these practitioners admitted to avoiding

the use of visualization in thumbnails. Instead, they opted

to incorporate photographic imagery whenever possible,

and that photos of people appear to drive more trafﬁc to

articles. Maps also appear to successfully drive readers to

articles. One practitioner reported that visualizations are

often regarded by readers as “cold, intimidating, or inacces-

sible,” and that while a visualization in isolation does not

communicate much in itself, they can sometimes be used

to accentuate a photographic or illustrated thumbnail. For

instance, a composite thumbnail can incorporate a photo

of a person with a semi-transparent visualization overlay.

This practitioner also avoided incorporating visualization

in thumbnails because they did not want to “give away too

much content” before the reader arrives at the article. Lastly,

we also learned that creating a thumbnail for an article is

often not the responsibility of the article author or visual-

ization designer. Instead, thumbnail design is designated

to designers and social media content producers who are

typically not involved in writing the article.

3.3 Visualization Thumbnail Deﬁnition and Goals

Based on the survey results and our interviews, we deﬁne

a visualization thumbnail as a thumbnail that includes one or

3. https://www.datavisualizationsociety.org

more visualizations. A thumbnail, in turn, is a small image

typically between 96 × 96 and 256 × 256 pixels in dimen-

sion designed as a preview of a larger one. Visualization

thumbnails are commonly designed to be informative to

allow readers to understand the gist of an article without

reading, as well as enticing readers to click the article for

further reading. This does not imply that a thumbnail must

be inviting and informative to be considered a visualiza-

tion thumbnail; only that if both of these design goals are

satisﬁed, the thumbnail can be considered a well-designed

visualization thumbnail.

Two perspectives of this deﬁnition exist, and each per-

spective leads to different visualization thumbnail design

goals. The goal of the professional visualization thumbnail

designer is to draw readers’ attention to the thumbnails and

to increase the click rate of the article per exposure. This

means that their design goals for a visualization thumbnail

goals are not much different from those of normal thumbnail

images. Therefore, conventional design goals, such as deter-

mining an image’s attractiveness [3] or visual aesthetics [38],

[45], [46], also apply to visualization thumbnail evaluation.

However, the goal of the consumer is to use the visu-

alization thumbnail to select articles that best match his or

her subjective criteria (e.g., preferences, interests, or inten-

tions). For these users, informativeness in thumbnails can

be the most important requirement that allows the readers

to quickly and accurately judge whether the thumbnail’s

associated article meets their selection criteria. This perspec-

tive leads to further design goals, such as informativeness,

relevance, interpretability, and straightforwardness [1], [2],

[6], [47], [48].

In an ideal scenario, the goals of producers and con-

sumers for visualization thumbnails are the same. However,

in our interviews, practitioners with extensive experience

asserted that there are cases in which they do not design

visualization thumbnails to simply meet readers’ needs

because they also need to consider user click counts.

In this paper, we seek to understand how to design vi-

sualization thumbnails to merge both of these perspectives.

More speciﬁcally, we are looking for a middle ground where

visualization thumbnails are both visually appealing—thus

inviting more clicks—as well as representative of the under-

lying article—thus ensuring that readers do not feel misled.

We believe our work to be the ﬁrst investigation of this topic.

Due to the limitations in the existing methods and scopes of

the study, the proposed deﬁnition should be considered as

a working deﬁnition and could evolve or be replaced with

the results of follow-up experiments and studies.

Though our interviews with practitioners were informal

and by no means exhaustive, we were encouraged to learn

about the lack of consensus in regards to guidelines for

visualization thumbnail design. We therefore remained cu-

rious about what makes visualization thumbnails effective,

or how their components contribute to whether readers ﬁnd

them to be inviting and informative. Conversely, we also

questioned which thumbnail components could be used

to mislead or misinform readers. For example, a photo of

United States special prosecutor Robert Mueller looking

down may deliver a feeling of disappointment to readers.

Considered alongside our survey results, these interviews

reinforced the need for further empirical study.

This article has been accepted for publication in IEEE Transactions on Visualization and Computer Graphics. This is the author's version which has not been fully edited and

content may change prior to final publication. Citation information: DOI 10.1109/TVCG.2023.3278304

Authorized licensed use limited to: Hong Kong University of Science and Technology. Downloaded on July 19,2023 at 07:46:22 UTC from IEEE Xplore. Restrictions apply.

IEEE TRANSACTIONS ON VISUALIZATION AND COMPUTER GRAPHICS 6

29 participants

GNRD

LeBron

Trump

Apple

vs.

Microsoft

Kavanaugh

53 participants

Highlight

52 participants

HRO

27 participants

Resized

Fig. 3. Four types of visualization thumbnails. 16 thumbnails of four types with four articles from different topics that are used in our experiment.

Columns ( , , , ) refer to thumbnail types, and rows (LeBron, Trump, Apple vs. Microsoft, Kavanaugh) refer to article types.

4 USER STUDY

Our survey and interviews revealed a lack of studies on

reader preferences for visualization thumbnail design. To

investigate these, we formulated two research questions

regarding (RQ1) what design types readers would most likely

click to read and why?, and (RQ2) what do readers think of

the components in visualization thumbnails? To answer these

questions, we conducted an online user study.

4.1 Visualization Thumbnail Design

There are three considerations for designing the thumbnails

for this study, including chart and editing types and com-

ponent combinations. In terms of chart type, we use the line

chart due to its popularity (Table 1) and interpretability [49].

Since we limit our discussion to the line chart thumbnails,

there remains a research question on extending the design

space for charts with different visualization components,

such as bar, bubble, or scatterplot, and maps. For the editing

type, we exclude cropped thumbnails because they could

cause unexpected issues, such as deception [32], [50], [51].

We initially considered a design with many components and

their combinations for RQ2, but soon found that there were

too many permutations (e.g., 262,144 combinations with 18

chart components), making a user study infeasible without

any guidelines for component selection.

We therefore design our visualization thumbnails based

on the chart component survey result [4], including the

basic chart components. We also include GNRDs (Graphics

Not Related to Data) [20], [25], [26], [27], [28], [29], HROs

(Human-Recognizable objects) [27], [36], titles, and visual

highlights with data labels [18], [52], [53], all of which affect

the appeal and interpretability of a thumbnail.

We select four articles for the study based on four con-

siderations; thumbnails should 1) be easy for a layperson to

read and understand; 2) include at least one visualization

that conveys the article’s main point; 3) have a visualization

for two data series that can be represented as icons or logos

to help readers distinguish two or more pieces of data;

and 4) contain the main character that can be displayed as

GNRDs. After a careful examination of candidate articles,

we ﬁnd four articles that satisfy our requirements. The

topics of the four articles pertain to sports star’s inﬂuence

on fan bases [54] (LeBron), a recent U.S. president’s approval

ratings [55] (Trump), a comparison of Apple and Microsoft

stock prices [56] (Apple vs. Microsoft), and a summary

judgment of a politician’s conﬁrmation [57] (Kavanaugh).

Among the four articles, the articles about Trump and Apple

vs. Microsoft contain only one visualization in each article.

The articles about LeBron and Kavanaugh include more

than two visualizations. When designing our visualization

thumbnails for the experiment, we referred to articles’

thumbnails or the visualization most discussed in the article.

Surveying visualization thumbnails with two experts in

the ﬁeld of design, we ﬁnd that thumbnails have many

design variables in the thumbnails that designers manip-

ulate to create strong impressions of brand identity and

invite more readers to click. These design variables include

colorful backgrounds [58], unique layouts (e.g., hiding chart

areas by a ﬁgure: [44]), and images of faces with stimulat-

ing facial expressions and gestures (e.g., a photograph of

Donald Trump: [59]). Due to its large decision space and

This article has been accepted for publication in IEEE Transactions on Visualization and Computer Graphics. This is the author's version which has not been fully edited and

content may change prior to final publication. Citation information: DOI 10.1109/TVCG.2023.3278304

Authorized licensed use limited to: Hong Kong University of Science and Technology. Downloaded on July 19,2023 at 07:46:22 UTC from IEEE Xplore. Restrictions apply.

IEEE TRANSACTIONS ON VISUALIZATION AND COMPUTER GRAPHICS 7

TABLE 3

Chart components. Identiﬁed for each thumbnail.

Y Axis Label

Y Grid Line

X Axis Label

Chart Title

Legend Implicit

Y Axis Title

X Axis

X Tick Marks

Legend Explicit

X Grid Lines

Data Label

Y Axis

Y Tick Marks

X Axis Title

Exp. text

GNRD

HRO

Modified

Cropped

Resized

SUM

Highlight

BASIC CHART COMP. ADDED CHART COMP.

Removed Added Remained Editing Type

GNRD

Highlight

HRO

Resized

possible bias, we decide to use ﬁxed values for the variables.

For example, we use the same color sets with original

visualization, the same layouts for the same thumbnail type,

and ﬁgures with neutral facial expressions.

Based on the considerations and articles, we then present

the designed visual thumbnails. We design 16 thumbnails of

four types, as shown in Figure 3. Table 3 summarizes how

we control the chart component combinations. To decide the

size of the visualization thumbnails, we survey the size of

visualization thumbnails from news organizations (supple-

mentary material https://osf.io/khgw2/). Among different

thumbnail sizes, we choose Pudding’s size (width: 343px,

height: 193px), because its width is close to an average (i.e.,

not extremely small or large), and it can sufﬁciently show all

components of the designed thumbnails in popular mobile

devices (e.g., iPhone X, display size: 375×812px).

thumbnails are the resized versions of the

visualizations from the original articles. The New York Times

and The Economist frequently use this design approach.

The charts in this type tend to contain many components,

including basic chart components, such as axis labels, im-

plicit legends, and a chart title, as many other thumbnails

do (e.g., [56]). There could be minor differences in axis

representations (e.g., tick marks) and the use of grid lines

in the original visualizations in the articles. For consistency

with thumbnails, we use the same style of tick marks

and gridlines across the thumbnails.

thumbnails consist of a photograph (GNRD) and

a simpliﬁed chart without any axis labels. GNRDs are

included in the thumbnails to invite readers, stimulating

their curiosity (e.g., the photography in this thumbnail of a

woman walking: [60]). They often imply information about

an article’s content by showing a person who is related to

the content. In general, the chart shows an overall trend

without axes, and words as implicit legends. This type

of thumbnail is frequently shown in FiveThirtyEight’s (e.g.,

[44]) and The Wall Street Journal’s graphics.

The main feature of thumbnails is the use

of a highlight block (e.g., red blocks in Fig. 3 )

to emphasize part of a visualization (e.g., [61]). Data

labels are added to emphasize a speciﬁc part of the data

that conveys the main point of the article (e.g., ‘46.0%’

and ‘39.8%’ of this thumbnail [55]). The Wall Street Journal

Graphics,FiveThirtyEight, and Bloomberg News frequently use

this thumbnail-design approach. In general, the visualiza-

tion in this type contains simpler axes than those in

thumbnails, excluding basic components such as axis labels

and gridlines. Reference lines can also have data labels,

which provide the information of the complete y-axis (e.g.,

the dotted vertical line and numbers in this thumbnail [55]).

thumbnails have recognizable objects, such as

logos or icons, on a plain line chart with grid lines as the

background. Their main feature is that both logos and icons

(e.g., icons of handcuffs [62]) directly refer to the data in the

visualization. (e.g., logos [54]). FiveThirtyEight and the Pew

Research Center frequently produce this type of thumbnail.

4.2 Workshop with Practitioners

To ensure the design quality of the designed thumbnails

and to collect feedback and expectations on the thumbnail

experiment, we held a workshop with 9 practitioners at a

data journalism conference. The practitioners were a data

scientist (ﬁve years of experience), three graphic designers

(six, four, and three years of experience, respectively), two

data journalists (ﬁve and four years of experience, respec-

tively), a computational journalist (three years of experi-

ence), and two interactive graphics developers (13 and three

years of experience, respectively). They all worked for data

journalism organizations.

As the workshop began, we described the four types

of visualization thumbnails and showed the 16 thumbnails

(Figure 3) produced with associated news articles. We asked

them to evaluate the quality of the thumbnails and their per-

sonal preferences for the thumbnail types. We also requested

that they report any possible issues that make a thumbnail

ineffective for use. Lastly, we asked them if they, based on

their domain experience, had comments, including possible

hypotheses or insights on the experiment.

Overall, we received positive feedback on the thumb-

nails; all practitioners agreed that there were no critical

issues that made the thumbnails look far different from

those they had created. For their preferences for thumbnails,

seven practitioners expressed their preference for , as

they are intuitive and eye-catching: “I can easily catch that

what I would read ‘some statistics of Donald Trump’ and ‘the sales

of stock prices of two companies’." One practitioner expressed

her preference for thumbnails, pointing out both their

interpretability and appeal: “News articles are similar to books

on a large shelf... The eye-catching design of a book is most

important. If the book cover is informative, it will do better."

Meanwhile, ﬁve practitioners stated they were interested in

our results because they assumed that readers could view

thumbnails as unfriendly, hard to read, and cluttered,

but they did not have evidence of this assumption.

To sum up, we found that the thumbnails we produced

are similar to what the practitioners create. We also con-

ﬁrmed that the thumbnails could be used in the user study

after revising them based on the collected design sugges-

tions (e.g., increasing font sizes, rounding line edges). The

practitioners were also curious about the readers’ evaluation

of the four types of thumbnails. Next, we describe how we

performed the user study with the designed thumbnails.

4.3 Study Procedure

For the experiment, we created a web page where read-

ers could browse news articles with thumbnails. For

the page layout, we referred to the graphics sections of

Bloomberg (https://www.bloomberg.com/graphics) and The

Wall Street Journal (https://graphics.wsj.com). For example,

the layout presented a news article thumbnail ﬁrst and then

placed a title and bylines below the thumbnail. The web

This article has been accepted for publication in IEEE Transactions on Visualization and Computer Graphics. This is the author's version which has not been fully edited and

content may change prior to final publication. Citation information: DOI 10.1109/TVCG.2023.3278304

Authorized licensed use limited to: Hong Kong University of Science and Technology. Downloaded on July 19,2023 at 07:46:22 UTC from IEEE Xplore. Restrictions apply.

IEEE TRANSACTIONS ON VISUALIZATION AND COMPUTER GRAPHICS 8

GNRD

Highlight

Resized

HRO

Fig. 4. Sample interface. Mobile page used for the experiment.

page listed four thumbnails at a time (Figure 4) and allowed

scrolling.

We conducted the user study with participants recruited

from Proliﬁc. We allowed the participants to participate

in the experiment only once. We also prevented duplicate

participation by matching their platform ID (Proliﬁc) to a

session key provided by the system to access our experi-

ment page. As participants entered the experiment website

published in Proliﬁc, they were asked to read the introduc-

tion and electronically agree with the consent form. They

were also provided access to the IRB approval document

(UNIISTIRB-19-33-A). Then, they were asked to ﬁll out a

demographic survey (e.g., gender, age, education, and fre-

quency of using a mobile device to read news articles). After

the survey, participants were asked to read instructions that

included the rules and restrictions for the experiment, such

as not to create new tabs or refresh the page.

After reading the instructions, the participants entered

the web page and clicked on an article that they wanted to

read. They were asked to write a reason for their choice (a

minimum of 100 characters). Participants were encouraged

to provide selection reasons for their choices and to avoid

writing reasons purely based on their personal preferences

on a particular topic/person/location (e.g., “I like to read

sports news more than other topics"). Finally, participants

were asked a consistency check question (e.g., “please click

the article that you chose”) and then end the session. The

latter was conducted to detect any crowdworkers who were

randomly clicking; since it was administered directly after

the original question, this was not a memory check, but

a consistency check. During the experiment, we recorded

the participants’ device types and sizes to ensure that only

mobile device users could participate. Participants did not

report any complaints about the difﬁculty or ambiguity of

the task instructions.

We recruited 237 participants and provided the same

compensation (£1.25, hourly wage rate: £9.86), regardless of

the result of the consistency check. However, we excluded

the data of 40 participants who did not show consistency in

their thumbnail selection. We also excluded the data of 36

participants who did not follow instructions for thumbnail

selection and chose articles based on their personal pref-

erences ( : 8, : 12, : 9, : 7). Specif-

ically, we excluded reasons that only mention personal

preferences without any opinion on the thumbnails. The

excluded reasons tend to only have expressions on the topic

(11 participants), article titles (4 participants), and charac-

ter/company (21 participants), such as “I chose this thumb-

nail because Apple is an incredibly popular company and it

is very concerning when they don’t do well ﬁnancially.” We

also excluded reasons where coders were unable to reach

unanimous agreement due to ambiguity. For example, we

rejected this reason: “The data seems to run parallel in a

way that the other data did not. The consistency in the data

is very compelling," because coders could not agree which

part of the thumbnail the statement is describing. When

participants clearly mentioned the thumbnail image as the

reason for their selection, we included the data and coded

it accordingly. A “clear mention” includes elements such

as charts (and chart components), companies, and main

characters in the article.

In total, we analyzed the responses of 161 participants

(33 years old, on average, σ=9). They had moderate familiar-

ity with visualizations (1 = not familiar at all, 7 = common;

average score was 4) and frequently read news articles

on mobile devices: more than seven times per week: 64

participants; more than ﬁve times: 42; more than three times:

40; and less than three times: 15. During the experiment, the

participants used mobile devices with an average height and

width of 775.84 px (σ=101.45) and 388.40 px (min=320px,

max=450px, σ=32.18), respectively.

Before proceeding with our analysis, we ﬁrst tested if

users differently choose thumbnails and found that par-

ticipants’ selections are signiﬁcantly different (p < .05).

We then investigated if there is the ordering effect; if the

participants choose thumbnails in a speciﬁc position (e.g.,

always the ﬁrst shown thumbnail). For this, we calculated

the probability of each position of thumbnails being selected

and divided this probability into four intervals from 0 to 1

(e.g., ﬁrst: ≈ 0.3, second: ≈ 0.53, third: ≈ 0.8, and last:

≈ 1). Using a random function that selects X from 0 to 1,

the number of thumbnails selected according to random X is

determined according to the probability interval described

above. For example, if random X is 0.47, it falls into the

second interval, and we suppose that the thumbnail in the

second position in thumbnail combinations is selected in our

simulation. For the simulation, we used 161 sequences that

are used in the user study. We conducted this simulation

2,000 times and obtained a result that the selections were

not affected by any positions of the presented thumbnails

(p < .05). We uploaded the simulation code and procedure

in the supplementary materials (https://osf.io/khgw2/).

We also found very little interaction between the articles

and thumbnail types (Cramer’s V = 0.058) in participant

choices.

5 RESULTS AND ANALYSIS

We performed statistical tests to determine the most popular

thumbnail type and the reasons stated by participants for

their choice (RQ1). Then we conducted a qualitative analysis

that investigated the readers’ opinions on the components of

the visualization thumbnails (RQ2).

This article has been accepted for publication in IEEE Transactions on Visualization and Computer Graphics. This is the author's version which has not been fully edited and

content may change prior to final publication. Citation information: DOI 10.1109/TVCG.2023.3278304

Authorized licensed use limited to: Hong Kong University of Science and Technology. Downloaded on July 19,2023 at 07:46:22 UTC from IEEE Xplore. Restrictions apply.

IEEE TRANSACTIONS ON VISUALIZATION AND COMPUTER GRAPHICS 9

TABLE 4

Visualization thumbnail comment codebook. Table summarizing the selection reasons with comment counts. The light blue and yellow text

backgrounds indicate the interpretable and inviting categories, respectively.

Theme Counts

Informative 10 (23%)

9 (21%)Interesting, eye-catching

Storytelling 6 (14%)

Less cluttered 6 (14%)

Stimulating curiosity 5 (12%)

Professional, unbiased 5 (12%)

Aesthetically pleasing 2 (4%)

Matching with an article title 0 (0%)

Theme Counts

15 (29%)Interesting, eye-catching

11 (21%)Stimulating curiosity

6 (11%)Less cluttered

5 (10%)Storytelling

4 (8%)Aesthetically pleasing

4 (8%)Matching with an article title

1 (2%)Professional, unbiased

Theme Counts

Informative 24 (27%)

18 (20%)Interesting, eye-catching

16 (18%)Storytelling

11 (12%)Stimulating curiosity

5 (6%)Professional, unbiased

10 (11%)Less cluttered

3 (3%)Matching with an article title

2 (2%)Aesthetically pleasing

Theme Counts

Interesting, eye-catching 20 (26%)

19 (25%)Informative

12 (16%)Less cluttered

8 (10%)Storytelling

5 (6%)Stimulating curiosity

6 (8%)Professional, unbiased

5 (6%)Aesthetically pleasing

2 (3%)Matching with an article title

6 (11%)Informative

Interpretable (51%) Inviting (49%) Interpretable (40%) Inviting (60%) Interpretable (60%) Inviting (40%) Interpretable (47%) Inviting (53%)

Resized

43 comments

GNRD

52 comments

Highlight

89 comments

HRO

77 comments

5.1 Readers’ Choice for Thumbnail Types (RQ1)

We start by studying thumbnail type preference.

5.1.1 Most Chosen Thumbnail Type

First, our results showed signiﬁcant preferences among the

four thumbnail types ( : 27, : 29, : 53, :

52) according to the chi-squared test (χ

(3, N=161, 14.975,

p=.0018).

To analyze the differences in users’ selection among

thumbnail types, we used a Thurstonian model [63], which

is a scaling method for converting people’s discrete selection

(e.g., a vote) into physical or mental concepts (e.g., partic-

ipants’ tendency levels on a chosen thumbnail compared

to others). The model assumes that individual rankings

are distributed around aggregate rankings. It also assumes

that a parameter exists based on the number of votes for

thumbnail , , , and and that the dis-

tribution of the values (i.e., preference) following Gaussian

can be estimated by probability. To quantitatively measure

the preference for thumbnail designs, we calculated the

posterior distributions of their choice using Monte Carlo

Markov-Chain (MCMC) [64], a popular method for approx-

imating a distribution through sampling.

To visualize the distribution, the frequency distribution

of the estimated thumbnail preference among the thumbnail

types, we ﬁrst set ’s frequency distribution as zero

(A=0, a vertical black line in Figure 5) and sample 2000

times from the user selection data using the MCMC method.

Figure 5 shows the distributions (y-axis) of the estimated

thumbnail preference (x-axis) of , , and ,

compared to . The preference scale means that the

standard deviation of the preference is 1. So the effect size

(difference of mean/standard deviation) corresponding to

Cohen’s d is the distance itself. For example, 0.1318 is the

effect size. If two distributions of and another type

overlap, this means the probability of a reader choosing

one of the two types is close to 50%, respectively. As the

overlapped part between the two distributions becomes

narrow, the probability of a reader choosing becomes

lower (i.e., a thumbnail types on the right side of the x-axis

has a higher preference probability compared to ).

To ﬁnd any signiﬁcant difference in the preference

distributions for the thumbnail types, we calculated the

probability that the differences between the sampled

values from any two types are positive. When we picked

-0.15

-0.1 -0.05 0 0.05 0.1 0.15 0.2 0.25 0.3

A = 0 = 0.0135

C = 0.1373

D = 0.1318

Significantly

Different (<.05)

Significantly Different

Preference

Frequency

Resized GNRD

HRO

Highlight

GNRD

Highlight

HRO

Fig. 5. Participant preferences. Posterior distribution of participants’

preferences for thumbnail types shows that readers are likely to prefer

and thumbnails more than those from and .

two thumbnails from and , we found that

there is a 99.5% probability that the thumbnail from

has a higher reader preference than that from

(GNRD<Highlig ht)

=99.5%). In the end, we found

that and thumbnails have signiﬁcantly

higher reader preference probability distributions than

those of and (P

(Resized<Highlight)

=99.7%,

(Resized<HRO)

=99.7%, P

(GNRD<Highlig ht )

=99.5%, and

(GNRD<HRO)

=99.6%).

5.1.2 Thumbnail Selection Reasons

To analyze the reasons for thumbnail selection, we used

a bottom-up approach to thematic analysis [65]. For this

analysis, the lead author of this work coded the selection

reasons by developing low-level themes. Then, two other

authors reviewed the themes and revised them together to

see if there is any disagreement with the proposed themes.

Once all themes were established, the three authors inde-

pendently coded the data again with the themes and later

aggregated their results. The ﬁnal agreement level of the

coders is 95.8% agreement (Fleiss’ Kappa = 0.26).

After building eight themes, we classiﬁed the themes

into two large categories, “inviting" and “interpretable",

which are the two main goals of visualization thumbnails

(Section 3.2). Table 5 summarizes the eight themes in two

categories, with exemplary comments. For inviting cate-

gories, we included themes related to drawing readers’

attention. For example, themes related to readers’ impres-

This article has been accepted for publication in IEEE Transactions on Visualization and Computer Graphics. This is the author's version which has not been fully edited and

content may change prior to final publication. Citation information: DOI 10.1109/TVCG.2023.3278304

Authorized licensed use limited to: Hong Kong University of Science and Technology. Downloaded on July 19,2023 at 07:46:22 UTC from IEEE Xplore. Restrictions apply.

IEEE TRANSACTIONS ON VISUALIZATION AND COMPUTER GRAPHICS 10

TABLE 5

Example comments. Comments have been grouped into eight themes

in the inviting and interpretable categories. Note that the study

participants are likely paying more attention to thumbnails than an

ordinary person would.

Category Theme Example

Informative (22%)

Storytelling (14%)

Interesting,

eye-catching (24%)

Stimulating

curiosity (13%)

Inviting

(48%)

Interpretable

(52%)

Aesthetically

pleasing (5%)

Matching with an

article title (3%)

“The graph seems very jittery and intense so it caught my eye..” (p160)

“I found the graph to be intriguing.” (p141)

“The fluctuations of the graph make me wonder about why the data is

so and what the article can say about it.” (P4)

“ the graph in the image made the most sense to me - the information

it was portraying seemed cleared.” (P73)

“It conveyed its Information in a simple and clear manner..” (p15)

“It is the least cluttered of the batch.” (p81)

Less cluttered

(13%)

“I found this graph to be the most visually compelling of the

thumbnails..” (p90)

Professional,

unbiased (6%)

“It lays the tone of professionalism.” (p29)

“It seemed to be the least misleading in its thumbnail graph.” (p80)

“It is the only graphic that seems to be telling a story right out of the

gate--at least one that I could start to make sense of..” (p13)

“The graphics and logo enforcing the articles headlines and made the

data more clear to understand.” (p115)

sions of a thumbnail, such as eye-catching, professional, or aes-

thetically pleasing, were classiﬁed into the inviting category.

For the interpretable category, we included themes related

to delivering articles’ content to readers. For example, less

cluttered belongs to the interpretable category because it

describes the thumbnails’ readability and visibility levels,

which are close to the method of information delivery. The

classiﬁcation resulted in 43 comments for , 52 com-

ments for , 89 comments for , and 77 comments

for . Each participant reported 1.6 reasons for selection,

on average.

Table 4 shows the thematic coding results, where we

found that participants differently selected thumbnails with

different reasons, according to a chi-squared test– [ :

(7, N=161, 14.116, p=.049], [ : χ

(7, N=161, 21.231,

p=.003], [

: χ

(7, N=161, 38.191, p=.0001], [ :

(7, N=161, 33.026, p=.0001] Speciﬁcally, readers chose

and thumbnails because they are informa-

tive, interesting, and telling stories. On the other hand, they

picked the highlights in thumbnails as they provide

“stimulating curiosity." Since the main difference between

and thumbnails are in axes, highlights, and

data labels, we conjecture that the components have mainly

inﬂuenced readers’ ﬁnal selection. Participants chose

and thumbnails due to their invitingness.

Reviewing all analysis results, we found the reasons that

readers prefer and thumbnails over and

thumbnails as follows. The low preference for

thumbnails was somewhat unexpected, as practitioners had

highly rated them due to their effective communication of

the articles’ main ideas without causing much visual clutter

(Section 4.2). However, this result implies that the ability

can be a double-edged sword, as it leads to low informa-

tiveness of the thumbnails, compared to other thumbnails,

as discussed in the following section.

5.2 Reader Opinions on Chart Components (RQ2)

We focus here on the axis information, data labels with high-

lights, and image components (GNRDs, HROs), as these are

the primary reasons for thumbnail selection and themes.

Abundant axes information may gain readers’ trust

but are rarely used for understanding data stories: One

characteristic of thumbnails is that they are the only

type of thumbnails that provide complete axes information

(e.g., axis labels and values). We ﬁrst found that participants

welcome the detailed axis information in thumbnails

for an accurate understanding of data stories. Nine of 27

participants chose the axis information as to their selection

reason. P25 stated—“The clearly visible labels on the x- and y-

axis, [...] effective scale helped me quickly understand it." We also

noticed that thumbnails with full axes information might

help readers perceive trustworthiness and professionalism

from the articles (two of 27), as P19 states—“I can say that the

thumbnail is my pick because it lays the tone of professionalism."

P25 also expressed a similar opinion—“[...] most convincing

and less likely to be skewed due to the clearly visible labels on the

x and y axis."

However, participants did not use the axis information

extensively for interpretation. Instead, 10 of 27 participants

focused on overall line chart trends (e.g., P4: “ﬂuctuations of

the graph make me wonder [...]", P5: “Apple is dropping [...]"),

which can be done quickly and easily. These interpretations

were not much different from the descriptions with

thumbnails, which were of interest, as they had an abstract

line with no axes (e.g., “The graph seems to show more people

have grown [...]”). These participants’ thumbnail reading be-

havior can be viewed as a natural effort for quickly catching

the main idea of thumbnails without a waste of time and

in-depth analysis. This illustrates why thumbnails

do not attract readers’ preference—they present too much

information for the reading behavior people tend to employ

for visualization thumbnails.

Data labels with highlights can provide a visual

guide for readers: Highlights [18], [52], [53] and reference

lines [33], [66], [67] have been effective methods for at-

tracting readers and enhancing readers’ understandability

of data stories with visualization. We conﬁrmed the effec-

tiveness of these techniques in thumbnails. For example,

reference lines (e.g., the black horizontal lines of

thumbnails in Figure 3) tend to make participants read data

stories in chronological order, working as anchors for the

data labels [33], [66], [67]. Nine of 53 participants described

how they understand the thumbnails with thumb-

nails, directly referring to the data labels—“I chose this article

(Article 2) [...] it was nearly a 50/50 split on January 23, 2017,

and on October 19, 2019 [...]," (P96). Such an increased under-

standing on the visualization could help participants better

infer the article’s context, as P76 describes—“[the thumbnail]

sets up the expectation the approval rating has dropped over the

last two years, so I basically know the premise of the article."

Combining multiple components can be a good strategy

to achieve both appeal and interpretability. For example, the

highlights combined with data labels can work as “a good

visual guide" (P109), where highlights point where readers

need to focus and data labels stress what readers should

understand. The red blocks for presenting time ranges and

data labels not only drew readers’ attention to speciﬁc data

points but also made readers curious about the article’s

context (e.g., P71: “The article shows a 41.2% approval rate

(in the red block), which is something I’d like to know more about,

This article has been accepted for publication in IEEE Transactions on Visualization and Computer Graphics. This is the author's version which has not been fully edited and

content may change prior to final publication. Citation information: DOI 10.1109/TVCG.2023.3278304

Authorized licensed use limited to: Hong Kong University of Science and Technology. Downloaded on July 19,2023 at 07:46:22 UTC from IEEE Xplore. Restrictions apply.

IEEE TRANSACTIONS ON VISUALIZATION AND COMPUTER GRAPHICS 11

i.e., the reasons").

GNRD and HRO, despite being considered eye-

grabbers, are associated with different keywords; GNRD

with “inspiration" and HRO with “informativeness": Im-

agery components, such as GNRDs and HROs, were effec-

tive in engaging readers and drawing their attention [18],

[20]. For thumbnails, 15 of 29 participants chose

GNRDs as their reason for the selection, because GNRDs

were “much more visually engaging" (P49) and “making the

thumbnail stand out more" (P41). Participants who chose

thumbnails also described HROs’ role as attention grabbers–

“The inclusion of sports logos caught my eye" (P132), “[...] logos

were recognizable and got my attention." (P152). We addition-

ally found a possibility that visualization thumbnails with

GNRDs could be more effective with a title in attracting

readers, which is the main access point for further read-

ing [2], [14], as P54 stated—“The photo caught my eye, which

leads to me reading the title."

Though both GNRDs and HROs attract readers, par-

ticipants’ interpretation of GNRDs and HROs may differ

in visualization thumbnails. In the study, participants with

HROs tend to intuitively understand the visualization, us-

ing HROs as “visual legends" (P136). For example, in the

thumbnail for Trump in Figure 3, the thumbs up

or thumbs down icons represent approval or disapproval,

which are easy to understand as they are consistent with

text labels (i.e., ‘approve’ and ‘disapprove’)—“Thumbs up

and thumbs down is a universally used graphic represen-

tation which was easy to understand for me." This supports

previous work stating that HROs can convey representative

meanings in the form of icons or logos but also create data

redundancy by conveying meaning combined with text [27].

For GNRDs, incorporating human photos in charts could

be inspiring but may lead to open interpretations. Three par-

ticipants (3 of 29) believed that GNRDs effectively informed

the article’s main character–“This one stood out because it

had a picture of a person mentioned in the article [...] (with the

GNRD) easier to think about because you don’t need to visualize

the person mentioned." We also observed that the meanings

of human photos might be accepted differently by readers.

For example, with the same thumbnail ( thumbnail for

the stock prices), P31 commented that “(The people in the

thumbnail) are selling Apple stock", while P35 reported that

“[...] they were talking about and what they determine is a bad

month for Apple [...]". We think this can be aligned with the

previous ﬁnding that thumbnails with human photographs

construe interpersonal and textual meanings [68].

We conjecture that informativeness of chart components

is another important factor that affects the ﬁnal selection

of thumbnails. For example, the participants may consider

thumbnails to be both attractive and informative, deliv-

ering the meaning of thumbnails in a straightforward man-

ner with text labels. GNRDs also attract readers’ attention,

but we ﬁnd the possibility that the participants could inter-

pret the meaning of GNRDs somewhat differently. However,

given the small number of participants mentioning the

difference between GNRDs and HROs, there may be more

as-yet unknown effects of GNRDs.

Inviting Interpretable

Axis

Components

Chart Title

GNRDs

HROs

Highlights

Data Labels

Highlights

HROs

Implicit Legends

Data Labels

Implicit Legends

Fig. 6. Visualization thumbnail roles. Interpretability and invitation.

6 DISCUSSION

We have shown that readers want visualization thumbnails

that not only keep their eyes focused on the thumbnail (i.e.,

inviting), but also convey the article’s main point (informa-

tive). Our ﬁndings indicate that thumbnail design can be

aided by understanding chart component roles.

6.1 Lessons Learned and Design Implications

Figure 6 summarizes the role of chart components for vi-

sualization thumbnails. We place GNRDs, HROs, and high-

lights in the inviting category, while we place axes, implicit

legends, and data labels in the interpretable category. We

also ﬁnd that data labels with highlights and HROs with

implicit legends can work for both inviting and interpretable

in the visualization thumbnails. We acknowledge that the

participants in our study are more likely to pay attention to

thumbnails than the general population. Here we provide

design suggestions that consider the roles of the chart com-

ponents and the combination of these components.

Lesson 1: Readers want easy-to-understand visualiza-

tion thumbnails. As thumbnails present stories, readers

expect to quickly and easily grasp the key information of

the story by viewing the thumbnails. As such, we think that

summarized and simple charts in thumbnails are preferable

compared to detailed charts, as shown in our results; readers

click thumbnails with brief summaries ( , 58) rather

than full charts for the articles ( , 27). We addition-

ally think that explicit and concise keywords can play an

important role in presenting the main point of the story,

while maintaining simple visualization thumbnails—“The

graph in the thumbnail is easier to look at. The large text makes it

a lot clearer what it’s supposed to be about at a glance compared

to the others,” as P72 states. Chart axes can aid readers in

understanding the thumbnail, but they may not be effective

if readers do not award them sufﬁcient attention.

Suggestion 1: Including descriptive text with a summarized

visualization can make a thumbnail more understandable. We

notice that chart titles and data labels help readers to

grasp an article’s key messages, making a thumbnail more

interpretable (blue part in Figure 6). Chart titles should

summarize the entire chart to help readers get a quick

and intuitive understanding on the thumbnails. Data labels

of numbers can be critical design points that help readers

access the detailed but concise information of the article.

Lesson 2: Readers want attractive visualization thumb-

nails. Conventionally, eye-catching has been an important

factor in thumbnail design [3], [9], [10], [68], [69]. We conﬁrm

in this work that readers also expect to see a thumbnail

This article has been accepted for publication in IEEE Transactions on Visualization and Computer Graphics. This is the author's version which has not been fully edited and

content may change prior to final publication. Citation information: DOI 10.1109/TVCG.2023.3278304

Authorized licensed use limited to: Hong Kong University of Science and Technology. Downloaded on July 19,2023 at 07:46:22 UTC from IEEE Xplore. Restrictions apply.

IEEE TRANSACTIONS ON VISUALIZATION AND COMPUTER GRAPHICS 12

that is attractive enough to draw their attention and invite

them for further clicks. In particular, we ﬁnd that readers

prefer thumbnails with attractive chart components. Many

readers ﬁrst catch images or highlights in thumbnails and

then become interested enough to explore the data and

read the article—“The red block made the thumbnail look more

entertaining to learn about" (P92).

Suggestion 2: Including highlights or visual embellishments

can make a visualization thumbnail engaging. Incorporating

highlights or images in thumbnails is a promising strategy

for drawing the reader’s attention to the thumbnail (i.e.,

the yellow part of Figure 6). Highlights make a part of the

chart stand out among other components and draw readers’

attention to the story. Visual embellishments such as GNRDs

and HROs can be also used for eye-catching purposes, but

we suggest that using them with cautions because it is

possible that they can mislead readers. For example, por-

trait photography (GNRDs) can bring unexpected meaning

when interpreted with given chart in a thumbnail.

6.2 Limitations and Open Research Areas

Although revealing readers’ thoughts on inviting and in-

terpretable visualization thumbnails throughout the user

study are the ﬁrst attempt at visualization thumbnail design,

there are limitations in this work. In this section, we discuss

the limitations of this work and present future research

questions beyond this work.

Exploring the role of chart components from quali-

tative analysis: We investigated readers’ opinion on chart

components and their roles (RQ2) providing other aspects

of visualization thumbnails that previous studies did not

point out. For example, we suggested the different roles of

HROs and GNRDs in visualization thumbnails which have

not been considered thus far. Though the obtained results

are meaningful in the sense that they are the ﬁrst responses

extracted from the readers, the result on the effectiveness

and interpretation of the chart components should not be

considered as deﬁnitive yet due to the limited number of

participants and extrapolations. We think a future study

could use the ﬁndings in this work in selecting chart compo-

nents to verify the speciﬁc role of the selected components

and their combinations. In this work, we focus on exploring

the role and effectiveness of different components in thumb-

nails. We think that a future study could investigate the role

of the components with consideration of topic, person, and

location perspectives.

Conﬁrming prior work: As the ﬁrst work for inves-

tigating readers’ preference and thoughts on visualization

thumbnail designs, we provided a series of ﬁndings, and

insights, and lessons from different analytical perspectives.

However, when compared to prior work, our ﬁndings may

be considered limited in that mostly conﬁrm existing ﬁnd-

ings. We think a future study may additionally validate

our hypotheses by focusing on the differences among other

types of thumbnails with different modiﬁcation strategies

based on our work.

Different visualization thumbnail sizes: Different read-

ers may have different thumbnail size preferences and fa-

miliarity. As our work is the ﬁrst study on visualization

thumbnail design, we used a single thumbnail size based

on our survey. Thus, the effect of thumbnail size remains

unanswered. A future study may investigate the possible

effects of different (smaller) visualization thumbnail sizes on

different computing platforms. This may implies there is a

space for testing various thumbnail sizes including different

size of components and fonts, as we described the roles

of effect of axes in Section 5.2. Our survey results show a

variety of sizes being used in online news media.

Presenting one message vs. summarizing a full story:

Designers can choose a message presentation strategy be-

tween conveying one fact/perspective or summarizing a

data story’s full narrative with a visualization thumbnail.

Among the visualization thumbnails used in the exper-

iments, we conjecture that the thumbnails seem

to support the story summary strategy by presenting a

baseline story with the time-series data and exposing the

main point with the red highlight and data label. This

can be related to picture superiority effect, in which image

content is more likely to lead to an increased understanding

of the material [70], [71], [72], [73], [74], [75], [76]. As we

did not consider the message presentation strategy in our

thumbnail design, there exist many questions on the matter.

For example, putting multiple or collage visualizations (e.g.,

bar+line, bar, and line) in one visualization thumbnail can

also be considered as a message presentation strategy, a

topic that can be further explored in future studies.

Design space generalization: In this work, we designed

and conducted a user study with visualization thumbnails

that present line charts inside them. As such, some insights

derived from this work may not be directly applicable to

other types of charts (e.g., maps) with different component

sizes, such as font sizes. For example, we described the

roles and effect of axes in Section 5.2, but further research

is needed to investigate whether the same effect occurs

with other chart types, such as bar chart, bubble chart,

or scatterplots. We used two editing types in this work—

resized and modiﬁed thumbnails—but a future study could

investigate other modiﬁcation strategies, such as cropped

thumbnails. Additional experiments can be performed for

investigating other possible impacts, such as the impacts of

interactions (e.g., scrolling) and surrounding colors. Lastly,

we used two time-series in this study because we wanted

to understand how readers understand implicit labels and

HROs in situations involving more than one time-series.

We were concerned that using more than two time-series

could lead to overly complicated experiments given the

lack of prior work or guidelines. Nevertheless, conducting

experiments with more than two time-series is an interesting

extension for future work, and could shed light on addi-

tional effective thumbnail designs.

Different highlighting techniques: To investigate the

impact of the presence of highlights in visualization thumb-

nails, we chose highlights that are frequently seen in the

thumbnail designs (e.g., red block, reference lines). How-

ever, other highlighting techniques could show better per-

formance. Examples include changing opacity or colors [30],

[33] and adding graphical annotations (e.g., circles) [33],

[77]. Future work should investigate the role of the highlight

techniques on readers’ perception.

Impact of stimulating portrait photography (GNRDs):

In this work, we used photos that have neutral impressions

This article has been accepted for publication in IEEE Transactions on Visualization and Computer Graphics. This is the author's version which has not been fully edited and

content may change prior to final publication. Citation information: DOI 10.1109/TVCG.2023.3278304

Authorized licensed use limited to: Hong Kong University of Science and Technology. Downloaded on July 19,2023 at 07:46:22 UTC from IEEE Xplore. Restrictions apply.

IEEE TRANSACTIONS ON VISUALIZATION AND COMPUTER GRAPHICS 13

and gestures in an effort to prevent any possible biases

from any other features related to charts (Section 4.1). With

the photos, we ﬁnd that portrait photography not only

catches readers’ attention but also tells a story when used

in visualization thumbnails, extending the previous works

[68], [69], [78], [79]. However, this result might not apply

to all other thumbnails, as there are many design choices

for photo selection in real-world thumbnails(e.g., a smiling

or frowning face). Future work could investigate the impact

of stimulating photos (GNRDs) with different features. For

example, we can study readers’ attention on various face

expressions of photography in visualization thumbnails,

compared to other components, such as headline texts.

Measuring a visualization thumbnail’s helpfulness af-

ter reading an article: In this work, we asked readers to

choose an article to read solely based on thumbnails. Thus,

we did not measure the perceived usefulness of visualiza-

tion thumbnails. Measuring the perceived helpfulness [2]

of visualization thumbnails is an important issue, as it

determines news organizations’ trust and reliability. For

example, after a reader reads an article, but the content

or conclusion of the article is not what is expected from

the thumbnail, the reader may skip thumbnails from the

news organizations. But what kind of factors or components

involves the usefulness of the thumbnails is not clear. Our

study indicates that GNRDs are one candidate that can affect

helpfulness of thumbnails, as they can be either informative

or misleading. As such, a future study needs to investigate

the perceived usefulness of visualization thumbnails and

the factors involved with the usefulness.

Impact of aesthetically-pleasing, artistic visualization

thumbnails: The visual aesthetics of visualization have

been discussed from many perspectives, including subjec-

tive impressions of visualization and design criteria [38],

[46]. In our study, the practitioners stated that they put much

effort on producing visually appealing thumbnails (sub-

section 3.2). We also observed many artistic visualization

thumbnails from news organizations. For example, FiveThir-

tyEight, the Wall Street Journal Graphics, and the Pudding tend

to adjust the color and contrast of photos (GNRDs) and

incorporate them into thumbnails together. However, due

to the larger design space and lack of clear criteria for de-

signing visually pleasing visualization thumbnails, we did

not discuss the topic in this work. For example, to design a

visually pleasing visualization thumbnail, one may need to

consider how to control biases from colorful backgrounds,

unique layouts, and face photos with different facial ex-

pressions and gestures. Future research could investigate

the aesthetics design space of visualization thumbnails and

answer what makes visualization thumbnails visually ap-

pealing to readers.

Honesty of responses: In the experiment, we followed

current best practice in crowdsourcing—we paid partici-

pants a sufﬁcient and fair amount of compensation and

did not give any systemic clue that could lead them to

certain answers. However, we cannot be sure of the hon-

esty of participant responses. On the other hand, we have

not found any indication that participants lied about their

answers. Furthermore, we believe that in our experimental

design, honesty is less effort than dishonesty for participants

because that means they don’t have to put effort into fabri-

cating dishonest answers and reasons.

Potential ordering effects: We used randomization

rather than counterbalancing, so there is a potential of an

ordering effect impacting our ﬁndings. To ﬁnd any pos-

sible ordering effect on thumbnail or article presentation

orders, we conducted a user thumbnail selection simulation

2,000 times as described in Section 4.3, where participants

choose thumbnails based on the selection probability in the

study. Thumbnails can be chosen in any position during the

simulations. Our ﬁndings from the simulation show that

participants were signiﬁcantly unlikely to be affected by

thumbnail positions.

7 CONCLUSION

We began this project by asking what makes thumbnails

for data stories inviting and interpretable. We surveyed

visualization thumbnails and had a series of interviews with

practitioners about the design of thumbnails for data-driven

stories. Based on our survey, we deﬁned the design types

of visualization thumbnails and conducted a user study

to determine the most appealing thumbnail design. Our

study results reveal a design space for thumbnails: a set of

thumbnail design guidelines that can be leveraged to attract

readers and help them understand the context of articles

associated with thumbnails. The results also indicate that

chart components are the keys in visualization thumbnails

to attract readers’ attention and enhance readers’ under-

standability of the visualization thumbnails. We also report

various thumbnail design strategies by effectively combin-

ing the chart components, such as a data summary with

highlights and data labels and a visual legend with text

labels and HROs. Ultimately, our study sheds light on an

uncharted design space for visualization thumbnail design

and toward automatically generating or recommending an

ideal set of visualization components to include in a thumb-

nail.

ACKNOWLEDGMENTS

This work was supported by the Korean National Re-

search Foundation (NRF) grant (No. 2021R1A2C1004542),

by a grant of the Korea Health Technology R&D Project

through the Korea Health Industry Development Institute

(KHIDI), funded by the Ministry of Health & Welfare,

Republic of Korea (grant number:HI22C0646), and by the In-

stitute of Information & Communications Technology Plan-

ning&Evaluation (IITP) grants (No.2020-0-01336–Artiﬁcial

Intelligence Graduate School Program, UNIST), funded by

the Korea government (MSIT).

REFERENCES

[1] Z. Li, S. Shi, and L. Zhang, “Improving relevance judgment of

web search results with image excerpts,” in Proceedings of the

International Conference on World Wide Web. New York, NY, USA:

ACM, 2008, pp. 21–30.

[2] A. Aula, R. M. Khan, Z. Guan, P. Fontes, and P. Hong, “A compari-

son of visual and textual page previews in judging the helpfulness

of web pages,” in Proceedings of the International Conference on World

Wide Web. New York, NY, USA: ACM, 2010, pp. 51–60.

[3] Y. Song, M. Redi, J. Vallmitjana, and A. Jaimes, “To click or not to

click: Automatic selection of beautiful thumbnails from videos,”

CoRR, vol. abs/1609.01388, 2016.

This article has been accepted for publication in IEEE Transactions on Visualization and Computer Graphics. This is the author's version which has not been fully edited and

content may change prior to final publication. Citation information: DOI 10.1109/TVCG.2023.3278304

Authorized licensed use limited to: Hong Kong University of Science and Technology. Downloaded on July 19,2023 at 07:46:22 UTC from IEEE Xplore. Restrictions apply.

IEEE TRANSACTIONS ON VISUALIZATION AND COMPUTER GRAPHICS 14

[4] H. Kim, J. Oh, Y. Han, S. Ko, M. Brehmer, and B. C. Kwon,

“Thumbnails for data stories: A survey of current practices,” in

Proceedings of the IEEE Visualization Conference. Piscataway, NJ,

USA: IEEE, 2019, pp. 116–120.

[5] Dziadosz, Susan and Chandrasekar, Raman, “Do thumbnail pre-

views help users make better relevance decisions about web search

results?” in Proceedings of the ACM Conference on Research and

Development in Information Retrieval. New York, NY, USA: ACM,

2002, pp. 365–366.

[6] A. Woodruff, A. Faulring, R. Rosenholtz, J. Morrison, and P. Pirolli,

“Using thumbnails to search the web,” in Proceedings of the ACM

Conference on Human Factors in Computing Systems. New York, NY,

USA: ACM, 2001, pp. 198–205.

[7] G. Robertson, M. Czerwinski, K. Larson, Robbins, D. C., D. Thiel,

and van Dantzich, Maarten, “Data mountain: Using spatial mem-

ory for document management,” in Proceedings of the ACM Sym-

posium on User Interface Software and Technology. New York, NY,

USA: ACM, 1998, pp. 153–162.

[8] A. Cockburn, C. Gutwin, and J. Alexander, “Faster document

navigation with space-ﬁlling thumbnails,” in Proceedings of the

ACM Conference on Human Factors in Computing Systems. New

York, NY, USA: ACM, 2006, pp. 1–10.

[9] J. Matejka, T. Grossman, and G. W. Fitzmaurice, “Swifter: im-

proved online video scrubbing,” in Proceedings of the ACM Con-

ference on Human Factors in Computing Systems. New York, NY,

USA: ACM, 2013, pp. 1159–1168.

[10] M. Topkara, S. Pan, J. C. Lai, A. Dirik, S. Wood, and J. Boston,

“"You’ve got video": increasing clickthrough when sharing enter-

prise video with email,” in Proceedings of the ACM Conference on

Human Factors in Computing Systems. New York, NY, USA: ACM,

2012, pp. 565–568.

[11] P. H. Nguyen, K. Xu, A. Bardill, B. Salman, K. Herd, and B. L. W.

Wong, “SenseMap: supporting browser-based online sensemaking

through analytic provenance,” in Proceedings of the IEEE Conference

on Visual Analytics Science and Technology. Piscataway, NJ, USA:

IEEE, 2016, pp. 91–100.

[12] V. Yoghourdjian, T. Dwyer, K. Klein, K. Marriott, and M. Wybrow,

“Graph thumbnails: Identifying and comparing multiple graphs at

a glance,” IEEE Transactions on Visualization and Computer Graphics,

vol. 24, no. 12, pp. 3081–3095, 2018.

[13] J. Heer, J. D. Mackinlay, C. Stolte, and M. Agrawala, “Graphical

histories for visualization: Supporting analysis, communication,

and evaluation,” IEEE Transactions on Visualization and Computer

Graphics, vol. 14, no. 6, pp. 1189–1196, 2008.

[14] S. Kaasten, S. Greenberg, and C. Edwards, “How people recognise

previously seen web pages from titles, URLs and thumbnails,” in

Proceedings of the British HCI Conference, 2002, pp. 247–266.

[15] B. Suh, H. Ling, B. B. Bederson, and D. W. Jacobs, “Automatic

thumbnail cropping and its effectiveness,” in Proceedings of the

ACM Symposium on User Interface Software and Technology. New

York, NY, USA: ACM, 2003, pp. 95–104.

[16] Z. Bylinskii, N. W. Kim, P. O’Donovan, S. Alsheikh, S. Madan,

H. Pﬁster, F. Durand, B. Russell, and A. Hertzmann, “Learning

visual importance for graphic designs and data visualizations,”

in Proceedings of the ACM Symposium on User Interface Software and

Technology. New York, NY, USA: ACM, 2017, pp. 57–69.

[17] M. Agrawala, W. Li, and F. Berthouzoz, “Design principles for

visual communication,” Communications of the ACM, vol. 54, no. 4,

pp. 60–69, 2011.

[18] J. Hullman, E. Adar, and P. Shah, “Beneﬁtting InfoVis with visual

difﬁculties,” IEEE Transactions on Visualization and Computer Graph-

ics, vol. 17, no. 12, pp. 2213–2222, 2011.

[19] E. R. Tufte, The Visual Display of Quantitative Information. Cheshire,

CT, USA: Graphics Press, 2001.

[20] S. Bateman, R. L. Mandryk, C. Gutwin, A. Genest, D. McDine, and

C. A. Brooks, “Useful junk?: the effects of visual embellishment

on comprehension and memorability of charts,” in Proceedings of

the ACM Conference on Human Factors in Computing Systems. New

York, NY, USA: ACM, 2010, pp. 2573–2582.

[21] D. A. Norman, “Simplicity is highly overrated,” Interactions,

vol. 14, no. 2, pp. 40–41, 2007.

[22] J. D. Kelly, “The data-ink ratio and accuracy of information de-

rived from newspaper graphs: An experimental test of the theory,”

in Proceedings of the Annual Meeting of the Association for Education

in Journalism and Mass Communication, 1988.

[23] D. Skau, L. Harrison, and R. Kosara, “An evaluation of the impact

of visual embellishments in bar charts,” Computer Graphics Forum,

vol. 34, no. 3, pp. 221–230, 2015.

[24] A. V. Moere, M. Tomitsch, C. Wimmer, B. Christoph, and

T. Grechenig, “Evaluating the effect of style in information visual-

ization,” IEEE Transactions on Visualization and Computer Graphics,

vol. 18, no. 12, pp. 2739–2748, 2012.

[25] R. Borgo, A. Abdul-Rahman, F. Mohamed, P. W. Grant, I. Reppa,

L. Floridi, and M. Chen, “An empirical study on using visual

embellishments in visualization,” IEEE Transactions on Visualization

and Computer Graphics, vol. 18, no. 12, pp. 2759–2768, 2012.

[26] M. A. Borkin, A. A. Vo, Z. Bylinskii, P. Isola, S. Sunkavalli, A. Oliva,

and H. Pﬁster, “What makes a visualization memorable?” IEEE

Transactions on Visualization and Computer Graphics, vol. 19, no. 12,

pp. 2306–2315, 2013.

[27] M. A. Borkin, Z. Bylinskii, N. W. Kim, C. M. Bainbridge, C. S. Yeh,

D. Borkin, H. Pﬁster, and A. Oliva, “Beyond memorability: Visu-

alization recognition and recall,” IEEE Transactions on Visualization

and Computer Graphics, vol. 22, no. 1, 2016.

[28] O. Inbar, N. Tractinsky, and J. Meyer, “Minimalism in information

visualization: attitudes towards maximizing the data-ink ratio,”

in Proceedings of ACM European Conference on Cognitive Ergonomics,

vol. 250. New York, NY, USA: ACM, 2007, pp. 185–188.

[29] S. Hill, B. Wray, and C. Sibona, “Minimalism in data visualization:

Perceptions of beauty, clarity, effectiveness, and simplicity,” in

Proceedings of the Conference on Information Systems Applied Research,

vol. 2167, 2017, p. 1508.

[30] D. Ren, M. Brehmer, B. Lee, T. Höllerer, and E. K. Choe, “Char-

tAccent: annotation for data-driven storytelling,” in Proceedings of

IEEE Paciﬁc Visualization Symposium. Piscataway, NJ, USA: IEEE,

2017, pp. 230–239.

[31] H.-K. Kong, Z. Liu, and K. Karahalios, “Frames and slants in titles

of visualizations on controversial topics,” in Proceedings of the ACM

Conference on Human Factors in Computing Systems. New York, NY,

USA: ACM, 2018, p. 438.

[32] A. V. Pandey, K. Rall, M. L. Satterthwaite, O. Nov, and E. Bertini,

“How deceptive are deceptive visualizations?: An empirical anal-

ysis of common distortion techniques,” in Proceedings of the ACM

Conference on Human Factors in Computing Systems. New York, NY,

USA: ACM, 2015, pp. 1469–1478.

[33] N. Kong and M. Agrawala, “Graphical overlays: Using layered

elements to aid chart reading,” IEEE Transactions on Visualization

and Computer Graphics, vol. 18, no. 12, pp. 2631–2638, 2012.

[34] C. D. Stolper, B. Lee, N. Henry Riche, and J. Stasko, “Data-driven

storytelling techniques: Analysis of a curated collection of visual

stories,” in Data-Driven Storytelling, N. Henry Riche, C. Hurter,

N. Diakopoulos, and S. Carpendale, Eds. A K Peters/CRC Press,

2018.

[35] R. Kosara, “Presentation-oriented visualization techniques,” IEEE

Computer Graphics and Applications, vol. 36, no. 1, pp. 80–85, 2016.

[36] S. Haroz, R. Kosara, and S. L. Franconeri, “Isotype visualiza-

tion: Working memory, performance, and engagement with pic-

tographs,” in Proceedings of the ACM Conference on Human Factors

in Computing Systems. New York, NY, USA: ACM, 2015, pp. 1191–

1200.

[37] L. Byrne, D. Angus, and J. Wiles, “Acquired codes of meaning

in data visualization and infographics: Beyond perceptual prim-

itives,” IEEE Transactions on Visualization and Computer Graphics,

vol. 22, no. 1, pp. 509–518, 2016.

[38] L. Harrison, K. Reinecke, and R. Chang, “Infographic aesthetics:

Designing for the ﬁrst impression,” in Proceedings of the ACM

Conference on Human Factors in Computing Systems. New York,

NY, USA: ACM, 2015, pp. 1187–1190.

[39] T. Andry, C. Hurter, F. Lambotte, P. Fastrez, and A. Telea, “In-

terpreting the effect of embellishment on chart visualizations,” in

Proceedings of the ACM Conference on Human Factors in Computing

Systems. New York, NY, USA: ACM, 2021, pp. 1–15.

[40] F. Amini, M. Brehmer, G. Bolduan, C. Elmer, and B. Wiederkehr,

“Evaluating data-driven stories & storytelling tools,” in Data-

Driven Storytelling, N. Henry Riche, C. Hurter, N. Diakopoulos,

and S. Carpendale, Eds. A K Peters/CRC Press, 2018.

[41] H. Kennedy, W. Weber, and M. Engebretsen, “Data visualization

and transparency in the news,” Data visualization in society, pp.

169–185, 2020.

[42] T. E. Team, “America is friendlier to foreigners than headlines sug-

gest,” The Economist, 2018, https://www.economist.com/sites/

default/ﬁles/20181215_WOC030_0.png.

This article has been accepted for publication in IEEE Transactions on Visualization and Computer Graphics. This is the author's version which has not been fully edited and

content may change prior to final publication. Citation information: DOI 10.1109/TVCG.2023.3278304

Authorized licensed use limited to: Hong Kong University of Science and Technology. Downloaded on July 19,2023 at 07:46:22 UTC from IEEE Xplore. Restrictions apply.

IEEE TRANSACTIONS ON VISUALIZATION AND COMPUTER GRAPHICS 15

[43] The New York Times’ Upshot, “Why Democrats’

gain was more impressive than it appears,” The

New York Times, 2018, https://static01.nyt.com/

images/2018/11/07/upshot/wave-result-1541613023801/

wave-result-1541613023801-mediumThreeByTwo252.png.

[44] O. Roeder, D. Mehta, and R. Dottle, “How cable

news covered Mueller in 2018,” FiveThirtyEight, 2018,

https://ﬁvethirtyeight.com/wp-content/uploads/2018/12/

YEARINMUELLER-4x3.png.

[45] A. V. Moere and H. Purchase, “On the role of design in information

visualization,” Information Visualization, vol. 10, no. 4, pp. 356–371,

2011.

[46] N. Cawthon and A. V. Moere, “The effect of aesthetic on the

usability of data visualization,” in Proceedings of the International

Conference on Information Visualization. Piscataway, NJ, USA: IEEE,

2007, pp. 637–648.

[47] H. Lam and P. Baudisch, “Summary thumbnails: readable

overviews for small screen web browsers,” in Proceedings of the

ACM Conference on Human Factors in Computing Systems. New

York, NY, USA: ACM, 2005, pp. 681–690.

[48] J. Teevan, E. Cutrell, D. Fisher, S. M. Drucker, G. Ramos, P. André,

and C. Hu, “Visual snippets: summarizing web pages for search

and revisitation,” in Proceedings of the ACM Conference on Human

Factors in Computing Systems. New York, NY, USA: ACM, 2009,

pp. 2023–2032.

[49] S. Lee, S.-H. Kim, and B. C. Kwon, “VLAT: Development of a

visualization literacy assessment test,” IEEE Transactions on Visual-

ization and Computer Graphics, vol. 23, no. 1, pp. 551–560, 2017.

[50] J. Ritchie, D. Wigdor, and F. Chevalier, “A lie reveals the truth:

Quasimodes for task-aligned data presentation,” in Proceedings of

the ACM Conference on Human Factors in Computing Systems. New

York, NY, USA: ACM, 2019, pp. 1–13.

[51] M. Correll, E. Bertini, and S. Franconeri, “Truncating the Y-axis:

Threat or menace?” in Proceedings of the ACM Conference on Human

Factors in Computing Systems. New York, NY, USA: ACM, 2020,

pp. 1–12.

[52] R. Kosara and J. Mackinlay, “Storytelling: The next step for visu-

alization,” Computer, vol. 46, no. 5, pp. 44–50, 2013.

[53] B. Lee, N. Henry Riche, P. Isenberg, and S. Carpendale, “More than

telling a story: Transforming data into visually shared stories,”

IEEE Computer Graphics and Applications, vol. 35, no. 5, pp. 84–90,

2015.

[54] O. Phillips, “What Reddit can tell us about NBA fan bases,”

FiveThirtyEight, 2018, https://ﬁvethirtyeight.com/features/

what-reddit-can-tell-us-about-nba-fan-bases/.

[55] S. Grocer, “How popular is Donald Trump?” FiveThirtyEight, 2019,

https://projects.ﬁvethirtyeight.com/trump-approval-ratings/

?ex_cid=rrpromo.

[56] A. Bycoffe, D. Mehta, and N. Silver, “Apple’s bad month is getting

worse, jeopardizing its status as most valuable company,” The

New York Times, 2018, https://www.nytimes.com/2018/11/27/

business/dealbook/apple-microsoft-stock.html.

[57] T. E. Team, “Brett Kavanaugh may have fared bet-

ter with senators than voters,” The Economist, 2018,

https://www.economist.com/graphic-detail/2018/09/28/

brett-kavanaugh-may-have-fared-better-with-senators-than-voters.

[58] A. Bycoffe, “Did Americans support removing Trump from

ofﬁce?” FiveThirtyEight, 2018, https://ﬁvethirtyeight.com/

wp-content/uploads/2019/10/IMPEACHMENT-TRACKER-4x3.

png.

[59] J. Eastwood, C. Jones, and J. Wolfe, “Trump, his chil-

dren, and 500+ potential conﬂicts of interest,” The Wall

Street Journal, 2017, https://si.wsj.net/public/resources/images/

OG-AJ911_201701_TOP_20170117190252.jpg.

[60] M. Koerth, “The pay gap is way too entrenched to be solved by

women alone,” FiveThirtyEight, 2018, https://ﬁvethirtyeight.com/

wp-content/uploads/2018/04/paygap-4x3.png.

[61] The New York Times, “Change in combined value

of Facebook, Amazon, Apple, Netﬂix and Alphabet,”

The New York Times, 2018, https://static01.nyt.com/

images/2018/11/21/business/dealbook/21db-techvalue1/

21db-techvalue1-mediumThreeByTwo440.png.

[62] A. Smith, “Majorities of Americans ﬁnd it unacceptable to use

algorithms to make decisions with real-world consequences

for humans,” The Pew Research Center, 2019, https://www.

pewresearch.org/internet/wp-content/uploads/sites/9/2018/

11/PI_2018.11.16_Attitudes-Toward-Algorithms_0-01.png.

[63] G. Yao and U. Böckenholt, “Bayesian estimation of Thurstonian

ranking models based on the Gibbs sampler,” British Journal of

Mathematical and Statistical Psychology, vol. 52, no. 1, pp. 79–92,

1999.

[64] J. Kruschke, Doing Bayesian Data Analysis: A Tutorial with R, JAGS,

and Stan. Academic Press, 2014.

[65] V. Braun and V. Clarke, “Using thematic analysis in psychology,”

Qualitative research in psychology, vol. 3, no. 2, pp. 77–101, 2006.

[66] D. Simkin and R. Hastie, “An information-processing analysis of

graph perception,” Journal of the American Statistical Association,

vol. 82, no. 398, pp. 454–465, 1987.

[67] J. Hullman, N. Diakopoulos, and E. Adar, “Contextiﬁer: automatic

generation of annotated stock visualizations,” in Proceedings of the

ACM Conference on Human Factors in Computing Systems. New

York, NY, USA: ACM, 2013, pp. 2707–2716.

[68] J. S. Knox, “Punctuating the home page: image as language in an

online newspaper,” Discourse & Communication, vol. 3, no. 2, pp.

145–172, 2009.

[69] ——, “Visual minimalism in hard news: thumbnail faces on the

smh online home page,” Social Semiotics, vol. 19, no. 2, pp. 165–

189, 2009.

[70] T. L. Childers and M. J. Houston, “Conditions for a picture-

superiority effect on consumer memory,” Journal of consumer re-

search, vol. 11, no. 2, pp. 643–654, 1984.

[71] R. N. Shepard, “Recognition memory for words, sentences, and

pictures,” Journal of verbal Learning and verbal Behavior, vol. 6, no. 1,

pp. 156–163, 1967.

[72] D. M. McBride and B. A. Dosher, “A comparison of conscious

and automatic memory processes for picture and word stimuli: A

process dissociation analysis,” Consciousness and cognition, vol. 11,

no. 3, pp. 423–460, 2002.

[73] A. J. Whitehouse, M. T. Maybery, and K. Durkin, “The develop-

ment of the picture-superiority effect,” British Journal of Develop-

mental Psychology, vol. 24, no. 4, pp. 767–773, 2006.

[74] B. A. Ally, C. A. Gold, and A. E. Budson, “The picture superiority

effect in patients with alzheimer’s disease and mild cognitive

impairment,” Neuropsychologia, vol. 47, no. 2, pp. 595–598, 2009.

[75] T. Curran and J. Doyle, “Picture superiority doubly dissociates the

erp correlates of recollection and familiarity,” Journal of Cognitive

Neuroscience, vol. 23, no. 5, pp. 1247–1262, 2011.

[76] C. Yangandul, S. Paryani, M. Le, and E. Jain, “How many words

is a picture worth? attention allocation on thumbnails versus title

text regions,” in Proceedings of the ACM Symposium on Eye Tracking

Research & Applications. New York, NY, USA: ACM, 2018, pp. 1–5.

[77] A. Satyanarayan and J. Heer, “Authoring narrative visualizations

with Ellipsis,” Computer Graphics Forum, vol. 33, no. 3, pp. 361–370,

2014.

[78] H. Caple and J. S. Knox, “Online news galleries, photojournalism

and the photo essay,” Visual Communication, vol. 11, no. 2, pp. 207–

236, 2012.

[79] A. W. Barrett and L. W. Barrington, “Bias in newspaper photo-

graph selection,” Political Research Quarterly, vol. 58, no. 4, pp.

609–618, 2005.

Hwiyeon Kim received her B.S. and M.S. degrees from UNIST (Ulsan

National Institute of Science and Technology) in Ulsan, South Korea in

2019 and 2021. Her research interests include data journalism and HCI.

Joohee Kim received her B.S. degree in Computer Science and Engi-

neering from UNIST (Ulsan National Institute of Science and Technol-

ogy) in Ulsan, South Korea. She is working toward her M.S. degree at

UNIST. Her research interests include data visualization and HCI.

Yunha Han received her M.S. in 2021 from UNIST (Ulsan National In-

stitute of Science and Technology) in Ulsan, South Korea. She is a data

engineer at NCSoft. Her research interests include data visualization

and HCI.

This article has been accepted for publication in IEEE Transactions on Visualization and Computer Graphics. This is the author's version which has not been fully edited and

content may change prior to final publication. Citation information: DOI 10.1109/TVCG.2023.3278304

Authorized licensed use limited to: Hong Kong University of Science and Technology. Downloaded on July 19,2023 at 07:46:22 UTC from IEEE Xplore. Restrictions apply.

IEEE TRANSACTIONS ON VISUALIZATION AND COMPUTER GRAPHICS 16

Hwajung Hong received the Ph.D. degree from Georgia Institute of

Technology in 2015. She is an assistant professor in the Department

of Industrial Design at KAIST. Her research lies at the intersection

of human-computer interaction and journalism. Her research interests

include the social implications of data and artiﬁcial intelligence.

Oh-Sang Kwon received the doctoral degree in psychological sciences

at Purdue University in 2009. He is an associate professor in the

department of biomedical engineering at UNIST (Ulsan National Insti-

tute of Science and Technology) in Ulsan, South Korea. His research

interests include human visual perception, perceptual/cognitive biases,

perception-action interaction, and perceptual learning.

Young-Woo Park received the Ph.D. degree in industrial design in 2014

from KAIST in Daejeon, South Korea. He is an associate professor of

design at UNIST (Ulsan National Institute of Science and Technology)

in Ulsan, South Korea. As a director of the Interactive Product Design

Laboratory, his research highlights the signiﬁcance of ’physical richness’

during interaction with technologies. He explores the potential of embod-

iment of tactile channel, shape-changing and data materialization as a

method for designing aesthetic interaction.

Niklas Elmqvist received the Ph.D. degree in

2006 from Chalmers University of Technology in

Göteborg, Sweden. He is a full professor in the

College of Information Studies at University of

Maryland, College Park in College Park, Mary-

land. He is also a member of the Institute for

Advanced Computer Studies (UMIACS) and for-

merly director of the Human-Computer Interac-

tion Laboratory (HCIL) at University of Maryland.

He is a senior member of the IEEE and the IEEE

Computer Society.

Sungahn Ko received the doctoral degree in

electrical and computer engineering in 2014

from Purdue University in West Lafayette, IN,

USA. He is an associate professor in the school

of Computer Science and Engineering at UNIST

(Ulsan National Institute of Science and Technol-

ogy) in Ulsan, South Korea. His research inter-

ests include visual analytics, information visual-

ization, and Human-Computer Interaction.

Bum Chul Kwon received his M.S. and Ph.D.

in Industrial Engineering from Purdue Univer-

sity in 2010 and 2013, and his B.S. in Systems

Engineering from University of Virginia in 2008.

He is a research scientist at IBM Research. His

research area includes visual analytics, data vi-

sualization, human-computer interaction, health-

care, and machine learning. Prior to joining IBM

Research, he worked as postdoctoral researcher

at University of Konstanz, Germany.

This article has been accepted for publication in IEEE Transactions on Visualization and Computer Graphics. This is the author's version which has not been fully edited and

content may change prior to final publication. Citation information: DOI 10.1109/TVCG.2023.3278304

Authorized licensed use limited to: Hong Kong University of Science and Technology. Downloaded on July 19,2023 at 07:46:22 UTC from IEEE Xplore. Restrictions apply.